RE: RAC partitioning idea bounce

From: Mark W. Farnham <mwf_at_rsiz.com>
Date: Fri, 6 Sep 2013 12:30:14 -0400
Message-ID: <040501ceab1e$5e37d910$1aa78b30$_at_rsiz.com>


First, see Tim Gorman's various descriptions of "Scaling to Infinity" to see if that method is a fit. IF it is a fit for this staging, that's probably the way to do it.

Second, David's issue of block mastering is real.

Third, instead of a hash, why not use instance_id for secondary partitioning. You would have to have a way to give a parameter for the instance id instead of the real current instance id in clean-up mode if some particular instance with a partially processed batch is down (and will continue to be down for too long).

This way, except in the broken instance case, you won't have dynamic block re-mastering and the partitioning matches both the current requirement for the processing job AND instance affinity. (You might experiment with having the partitions in tablespaces organized by instance to enhance the lessening of opportunity for block remastering. That may be over engineering.)

Someone's probably got a really slick way to do this that I haven't thought of, and maybe the "Scaling to Infinity" partitioning swapping deal is useful.

mwf

-----Original Message-----

From: oracle-l-bounce_at_freelists.org [mailto:oracle-l-bounce_at_freelists.org] On Behalf Of David Fitzjarrell
Sent: Friday, September 06, 2013 12:09 PM To: Jed_Walker_at_cable.comcast.com; oracle-l_at_freelists.org Subject: Re: RAC partitioning idea bounce

Your logic is sound, as I see it, but the RAC performance issue may not be caused solely by the segment contention.  Depending on the volume of data inserted running multi-node inserts into a table may cause Oracle to re-master the blocks from the current node to the busiest node  at the time, and if that activity occurs on different nodes in succession the constant re-mastering may create undue waits in addition to the block contention.  Using a single node prevents this re-mastering.  
My two cents.
David Fitzjarrell  



 From: "Walker, Jed S" <Jed_Walker_at_cable.comcast.com> To: "oracle-l_at_freelists.org" <oracle-l_at_freelists.org> Sent: Friday, September 6, 2013 9:55 AM
Subject: RAC partitioning idea bounce   

Hi All,
I have a system on 11.2.0.3 4-node RAC that I'm trying to help with. They have a process that inserts rows into a table. A database job then takes those rows and processes them into another table. There is a service configured on the RAC so that the inserting applications only connect to one RAC instance because they say the contention is too great when running across all nodes. I'm figuring the contention is due to all 4 nodes trying to insert into the same segment.

So, I thought, what if I partition based on the sole index (UK) on the table using Hash partitioning - to spread the inserts into multiple segments. The catch is that the table is partitioned already on a special partition_value column that is used by the processing job so it can process one partition of data at a time and then truncate the partition.  So, if I re-partition on the UK column then the processing job can't truncate the partitions any more. I don't have the option of changing all the processing code.

So, I think this is my best option and I'm looking for the insight of anyone with lots of partitioning experience. Recreate the table with partitioning on the partition_value column so that the job can do run the way it does currently, and then sub-partition on the UK column using a hash partition so that inserts will be distributed into many segments to reduce the multi-node contention.

Am I thinking through this correctly?

If I haven't provided enough info please ask for it.

Thanks,

Jed

--

http://www.freelists.org/webpage/oracle-l

--

http://www.freelists.org/webpage/oracle-l

--

http://www.freelists.org/webpage/oracle-l Received on Fri Sep 06 2013 - 18:30:14 CEST

Original text of this message